Related
I have 100 elements. Each element has 4 features A,B,C,D. Each feature is an integer.
I want to select 2 elements for each feature, so that I have selected a total of 8 distinct elements. I want to maximize the sum of the 8 selected features A,A,B,B,C,C,D,D.
A greedy algorithm would be to select the 2 elements with highest A, then the two elements with highest B among the remaining elements, etc. However, this might not be optimal, because the elements that have highest A could also have a much higher B.
Do we have an algorithm to solve such a problem optimally?
This can be solved as a minimum cost flow problem. In particular, this is an Assignment problem
First of all, see that we only need the 8 best elements of each features, meaning 32 elements maximum. It should even be possible to cut the search space further (as if the 2 best elements of A is not one of the 6 best elements of any other feature, we can already assigne those 2 elements to A, and each other feature only needs to look at the first 6 best elements. If it's not clear why, I'll try to explain further).
Then we make the vertices S,T and Fa,Fb,Fc,Fd and E1,E2,...E32, with the following edge :
for each vertex Fx, an edge from S to Fx with maximum flow 2 and a weight of 0 (as we want 2 element for each feature)
for each vertex Ei, an edge from Fx to Ei if Ei is one of the top elements of feature x, with maximum flow 1 and weight equal to the negative value of feature x of Ei. (negative because the algorithm will find the minimum cost)
for each vertex Ei, an edge from Ei to T, with maximum flow 1 and weight 0. (as each element can only be selected once)
I'm not sure if this is the best way, but It should work.
As suggested per #AloisChristen, this can be written as an assignment problem:
On the one side, we select the 8 best elements for each feature; that's 32 elements or less, since one element might be in the best 8 for more than one feature;
On the other side, we put 8 seats A,A,B,B,C,C,D,D
Solve the resulting assignment problem.
Here the problem is solved using scipy's linear_sum_assignment optimization function:
from numpy.random import randint
from numpy import argpartition, unique, concatenate
from scipy.optimize import linear_sum_assignment
# PARAMETERS
n_elements = 100
n_features = 4
n_per_feature = 2
# RANDOM DATA
data = randint(0, 21, (n_elements, n_features)) # random data with integer features between 0 and 20 included
# SELECT BEST 8 CANDIDATES FOR EACH FEATURE
n_selected = n_features * n_per_feature
n_candidates = n_selected * n_features
idx = argpartition(data, range(-n_candidates, 0), axis=0)
idx = unique(idx[-n_selected:].ravel())
candidates = data[idx]
n_candidates = candidates.shape[0]
# SOLVE ASSIGNMENT PROBLEM
cost_matrix = -concatenate((candidates,candidates), axis=1) # 8 columns in order ABCDABCD
element_idx, seat_idx = linear_sum_assignment(cost_matrix)
score = -cost_matrix[element_idx, seat_idx].sum()
# DISPLAY RESULTS
print('SUM OF SELECTED FEATURES: {}'.format(score))
for e,s in zip(element_idx, seat_idx):
print('{:2d}'.format(idx[e]),
'ABCDABCD'[s],
-cost_matrix[e,s],
data[idx[e]])
Output:
SUM OF SELECTED FEATURES: 160
3 B 20 [ 5 20 14 11]
4 A 20 [20 9 3 12]
6 C 20 [ 3 3 20 8]
10 A 20 [20 10 9 9]
13 C 20 [16 12 20 18]
23 D 20 [ 6 10 4 20]
24 B 20 [ 5 20 6 8]
27 D 20 [20 13 19 20]
I am searching for a method to create, in a fast way a random matrix A with the follwing properties:
A = transpose(A)
A(i,i) = 0 for all i
A(i,j) >= 0 for all i, j
sum(A) =~ degree; the sum of rows are randomly distributed by a distribution I want to specify (here =~ means approximate equality).
The distribution degree comes from a matrix orig, specifically degree=sum(orig), thus I know that matrices with this distribution exist.
For example: orig=[0 12 7 5; 12 0 1 9; 7 1 0 3; 5 9 3 0]
orig =
0 12 7 5
12 0 1 9
7 1 0 3
5 9 3 0
sum(orig)=[24 22 11 17];
Now one possible matrix A=[0 11 5 8, 11 0 4 7, 5 4 0 2, 8 7 2 0] is
A =
0 11 5 8
11 0 4 7
5 4 0 2
8 7 2 0
with sum(A)=[24 22 11 17].
I am trying this for quite some time, but unfortunatly my two ideas didn't work:
version 1:
I switch Nswitch times two random elements: A(k1,k3)--; A(k1,k4)++; A(k2,k3)++; A(k2,k4)--; (the transposed elements aswell).
Unfortunatly, Nswitch = log(E)*E (with E=sum(sum(nn))) in order that the Matrices are very uncorrelated. As my E > 5.000.000, this is not feasible (in particular, as I need at least 10 of such matrices).
version 2:
I create the matrix according to the distribution from scratch. The idea is, to fill every row i with degree(i) numbers, based on the distribution of degree:
nn=orig;
nnR=zeros(size(nn));
for i=1:length(nn)
degree=sum(nn);
howmany=degree(i);
degree(i)=0;
full=rld_cumsum(degree,1:length(degree));
rr=randi(length(full),[1,howmany]);
ff=full(rr);
xx=i*ones([1,length(ff)]);
nnR = nnR + accumarray([xx(:),ff(:)],1,size(nnR));
end
A=nnR;
However, while sum(A')=degree, sum(A) systematically deviates from degree, and I am not able to find the reason for that.
Small deviations from degree are fine of course, but there seem to be systmatical deviations in particulat of the matrices contain in some places large numbers.
I would be very happy if somebody could either show me a fast method for version1, or a reason for the systematic deviation of the distribution in version 2, or a method to create such matrices in a different way. Thank you!
Edit:
This is the problem in matsmath's proposed solution:
Imagine you have the matrix:
orig =
0 12 3 1
12 0 1 9
3 1 0 3
1 9 3 0
with r(i)=[16 22 7 13].
Step 1: r(1)=16, my random integer partition is p(i)=[0 7 3 6].
Step 2: Check that all p(i)<=r(i), which is the case.
Step 3:
My random matrix starts looks like
A =
0 7 3 6
7 0 . .
3 . 0 .
6 . . 0
with the new row sum vector rnew=[r(2)-p(2),...,r(n)-p(n)]=[15 4 7]
Second iteration (here the problem occures):
Step 1: rnew(1)=15, my random integer partition is p(i)=[0 A B]: rnew(1)=15=A+B.
Step 2: Check that all p(i)<=rnew(i), which gives A<=4, B<=7. So A+B<=11, but A+B has to be 15. contradiction :-/
Edit2:
This is the code representing (to the best of my knowledge) the solution posted by David Eisenstat:
orig=[0 12 3 1; 12 0 1 9; 3 1 0 3; 1 9 3 0];
w=[2.2406 4.6334 0.8174 1.6902];
xfull=zeros(4);
for ii=1:1000
rndmat=[poissrnd(w(1),1,4); poissrnd(w(2),1,4); poissrnd(w(3),1,4); poissrnd(w(4),1,4)];
kkk=rndmat.*(ones(4)-eye(4)); % remove diagonal
hhh=sum(sum(orig))/sum(sum(kkk))*kkk; % normalisation
xfull=xfull+hhh;
end
xf=xfull/ii;
disp(sum(orig)); % gives [16 22 7 13]
disp(sum(xf)); % gives [14.8337 9.6171 18.0627 15.4865] (obvious systematic problem)
disp(sum(xf')) % gives [13.5230 28.8452 4.9635 10.6683] (which is also systematically different from [16, 22, 7, 13]
Since it's enough to approximately preserve the degree sequence, let me propose a random distribution where each entry above the diagonal is chosen according to a Poisson distribution. My intuition is that we want to find weights w_i such that the i,j entry for i != j has mean w_i*w_j (all of the diagonal entries are zero). This gives us a nonlinear system of equations:
for all i, (sum_{j != i} w_i*w_j) = d_i,
where d_i is the degree of i. Equivalently,
for all i, w_i * (sum_j w_j) - w_i^2 = d_i.
The latter can be solved by applying Newton's method as described below from a starting solution of w_i = d_i / sqrt(sum_j d_j).
Once we have the w_is, we can sample repeatedly using poissrnd to generate samples of multiple Poisson distributions at once.
(If I have time, I'll try implementing this in numpy.)
The Jacobian matrix of the equation system for a 4 by 4 problem is
(w_2 + w_3 + w_4) w_1 w_1 w_1
w_2 (w_1 + w_3 + w_4) w_2 w_2
w_3 w_3 (w_1 + w_2 + w_4) w_3
w_4 w_4 w_4 (w_1 + w_2 + w_3).
In general, let A be a diagonal matrix where A_{i,i} = sum_j w_j - 2*w_i. Let u = [w_1, ..., w_n]' and v = [1, ..., 1]'. The Jacobian can be written J = A + u*v'. The inverse is given by the Sherman--Morrison formula
A^-1*u*v'*A^-1
J^-1 = (A + u*v')^-1 = A^-1 - -------------- .
1 + v'*A^-1*u
For the Newton step, we need to compute J^-1*y for some given y. This can be done straightforwardly in time O(n) using the above equation. I'll add more detail when I get the chance.
First approach (based on version2)
Let your row sum vector given by the matrix orig [r(1),r(2),...,r(n)].
Step 1. Take a random integer partition of the integer r(1) into exactly n-1 parts, say p(2), p(3), ..., p(n)
Step 2. Check if p(i)<=r(i) for all i=2...n. If not, go to Step 1.
Step 3. Fill out your random matrix first row and colum by the entries 0, p(2), ... , p(n), and consider the new row sum vector [r(2)-p(2),...,r(n)-p(n)].
Repeat these steps with a matrix of order n-1.
The point is, that you randomize one row at a time, and reduce the problem to searching for a matrix of size one less.
As pointed out by OP in the comment, this naive algorithm fails. The reason is that the matrices in question have a further necessary condition on their entries as follows:
FACT:
If A is an orig matrix with row sums [r(1), r(2), ..., r(n)] then necessarily for every i=1..n it holds that r(i)<=-r(i)+sum(r(j),j=1..n).
That is, any row sum, say the ith, r(i), is necessarily at most as big as the sum of the other row sums (not including r(i)).
In light of this, a revised algorithm is possible. Note that in Step 2b. we check if the new row sum vector has the property discussed above.
Step 1. Take a random integer partition of the integer r(1) into exactly n-1 parts, say p(2), p(3), ..., p(n)
Step 2a. Check if p(i)<=r(i) for all i=2...n. If not, go to Step 1.
Step 2b. Check if r(i)-p(i)<=-r(i)+p(i)+sum(r(j)-p(j),j=2..n) for all i=2..n. If not, go to Step 1.
Step 3. Fill out your random matrix first row and colum by the entries 0, p(2), ... , p(n), and consider the new row sum vector [r(2)-p(2),...,r(n)-p(n)].
Second approach (based on version1)
I am not sure if this approach gives you random matrices, but it certainly gives you different matrices.
The idea here is to change some parts of your orig matrix locally, in a way which maintains all of its properties.
You should look for a random 2x2 submatrix below the main diagonal which contains strictly positive entries, like [[a,b],[c,d]] and perturbe its contents by a random value r to [[a+r,b-r],[c-r,d+r]]. You make the same change above the main diagonal too, to keep your new matrix symmetric. Here the point is that the changes within the entries "cancel" each other out.
Of course, r should be chosen in a way such that b-r>=0 and c-r>=0.
You can pursue this idea to modify larger submatrices too. For example, you might choose 3 random row coordinates r1, r2, r2 and 3 random column coordinates c1, c2, and c3 and then make changes in your orig matrix at the 9 positions (ri,cj) as follows: you change your 3x3 submatrix [[a b c],[d e f], [g h i]] to [[a-r b+r c] [d+r e f-r], [g h-r i+r]]. You do the same at the transposed places. Again, the random value r must be chosen in a way so that a-r>=0 and f-r>=0 and h-r>=0. Moreover, c1 and r1, and c3 and r3 must be distinct as you can't change the 0 entries in the main diagonal of the matrix orig.
You can repeat such things over and over again, say 100 times, until you find something which looks random. Note that this idea uses the fact that you have existing knowledge of a solution, this is the matrix orig, while the first approach does not use such knowledge at all.
The cycle leader iteration algorithm is an algorithm for shuffling an array by moving all even-numbered entries to the front and all odd-numbered entries to the back while preserving their relative order. For example, given this input:
a 1 b 2 c 3 d 4 e 5
the output would be
a b c d e 1 2 3 4 5
This algorithm runs in O(n) time and uses only O(1) space.
One unusual detail of the algorithm is that it works by splitting the array up into blocks of size 3k+1. Apparently this is critical for the algorithm to work correctly, but I have no idea why this is.
Why is the choice of 3k + 1 necessary in the algorithm?
Thanks!
This is going to be a long answer. The answer to your question isn't simple and requires some number theory to fully answer. I've spent about half a day working through the algorithm and I now have a good answer, but I'm not sure I can describe it succinctly.
The short version:
Breaking the input into blocks of size 3k + 1 essentially breaks the input apart into blocks of size 3k - 1 surrounded by two elements that do not end up moving.
The remaining 3k - 1 elements in the block move according to an interesting pattern: each element moves to the position given by dividing the index by two modulo 3k.
This particular motion pattern is connected to a concept from number theory and group theory called primitive roots.
Because the number two is a primitive root modulo 3k, beginning with the numbers 1, 3, 9, 27, etc. and running the pattern is guaranteed to cycle through all the elements of the array exactly once and put them into the proper place.
This pattern is highly dependent on the fact that 2 is a primitive root of 3k for any k ≥ 1. Changing the size of the array to another value will almost certainly break this because the wrong property is preserved.
The Long Version
To present this answer, I'm going to proceed in steps. First, I'm going to introduce cycle decompositions as a motivation for an algorithm that will efficiently shuffle the elements around in the right order, subject to an important caveat. Next, I'm going to point out an interesting property of how the elements happen to move around in the array when you apply this permutation. Then, I'll connect this to a number-theoretic concept called primitive roots to explain the challenges involved in implementing this algorithm correctly. Finally, I'll explain why this leads to the choice of 3k + 1 as the block size.
Cycle Decompositions
Let's suppose that you have an array A and a permutation of the elements of that array. Following the standard mathematical notation, we'll denote the permutation of that array as σ(A). We can line the initial array A up on top of the permuted array σ(A) to get a sense for where every element ended up. For example, here's an array and one of its permutations:
A 0 1 2 3 4
σ(A) 2 3 0 4 1
One way that we can describe a permutation is just to list off the new elements inside that permutation. However, from an algorithmic perspective, it's often more helpful to represent the permutation as a cycle decomposition, a way of writing out a permutation by showing how to form that permutation by beginning with the initial array and then cyclically permuting some of its elements.
Take a look at the above permutation. First, look at where the 0 ended up. In σ(A), the element 0 ended up taking the place of where the element 2 used to be. In turn, the element 2 ended up taking the place of where the element 0 used to be. We denote this by writing (0 2), indicating that 0 should go where 2 used to be, and 2 should go were 0 used to be.
Now, look at the element 1. The element 1 ended up where 4 used to be. The number 4 then ended up where 3 used to be, and the element 3 ended up where 1 used to be. We denote this by writing (1 4 3), that 1 should go where 4 used to be, that 4 should go where 3 used to be, and that 3 should go where 1 used to be.
Combining these together, we can represent the overall permutation of the above elements as (0 2)(1 4 3) - we should swap 0 and 2, then cyclically permute 1, 4, and 3. If we do that starting with the initial array, we'll end up at the permuted array that we want.
Cycle decompositions are extremely useful for permuting arrays in place because it's possible to permute any individual cycle in O(C) time and O(1) auxiliary space, where C is the number of elements in the cycle. For example, suppose that you have a cycle (1 6 8 4 2). You can permute the elements in the cycle with code like this:
int[] cycle = {1, 6, 8, 4, 2};
int temp = array[cycle[0]];
for (int i = 1; i < cycle.length; i++) {
swap(temp, array[cycle[i]]);
}
array[cycle[0]] = temp;
This works by just swapping everything around until everything comes to rest. Aside from the space usage required to store the cycle itself, it only needs O(1) auxiliary storage space.
In general, if you want to design an algorithm that applies a particular permutation to an array of elements, you can usually do so by using cycle decompositions. The general algorithm is the following:
for (each cycle in the cycle decomposition algorithm) {
apply the above algorithm to cycle those elements;
}
The overall time and space complexity for this algorithm depends on the following:
How quickly can we determine the cycle decomposition we want?
How efficiently can we store that cycle decomposition in memory?
To get an O(n)-time, O(1)-space algorithm for the problem at hand, we're going to show that there's a way to determine the cycle decomposition in O(1) time and space. Since everything will get moved exactly once, the overall runtime will be O(n) and the overall space complexity will be O(1). It's not easy to get there, as you'll see, but then again, it's not awful either.
The Permutation Structure
The overarching goal of this problem is to take an array of 2n elements and shuffle it so that even-positioned elements end up at the front of the array and odd-positioned elements end up at the end of the array. Let's suppose for now that we have 14 elements, like this:
0 1 2 3 4 5 6 7 8 9 10 11 12 13
We want to shuffle the elements so that they come out like this:
0 2 4 6 8 10 12 1 3 5 7 9 11 13
There are a couple of useful observations we can have about the way that this permutation arises. First, notice that the first element does not move in this permutation, because even-indexed elements are supposed to show up in the front of the array and it's the first even-indexed element. Next, notice that the last element does not move in this permutation, because odd-indexed elements are supposed to end up at the back of the array and it's the last odd-indexed element.
These two observations, put together, means that if we want to permute the elements of the array in the desired fashion, we actually only need to permute the subarray consisting of the overall array with the first and last elements dropped off. Therefore, going forward, we are purely going to focus on the problem of permuting the middle elements. If we can solve that problem, then we've solved the overall problem.
Now, let's look at just the middle elements of the array. From our above example, that means that we're going to start with an array like this one:
Element 1 2 3 4 5 6 7 8 9 10 11 12
Index 1 2 3 4 5 6 7 8 9 10 11 12
We want to get the array to look like this:
Element 2 4 6 8 10 12 1 3 5 7 9 11
Index 1 2 3 4 5 6 7 8 9 10 11 12
Because this array was formed by taking a 0-indexed array and chopping off the very first and very last element, we can treat this as a one-indexed array. That's going to be critically important going forward, so be sure to keep that in mind.
So how exactly can we go about generating this permutation? Well, for starters, it doesn't hurt to take a look at each element and to try to figure out where it began and where it ended up. If we do so, we can write things out like this:
The element at position 1 ended up at position 7.
The element at position 2 ended up at position 1.
The element at position 3 ended up at position 8.
The element at position 4 ended up at position 2.
The element at position 5 ended up at position 9.
The element at position 6 ended up at position 3.
The element at position 7 ended up at position 10.
The element at position 8 ended up at position 4.
The element at position 9 ended up at position 11.
The element at position 10 ended up at position 5.
The element at position 11 ended up at position 12.
The element at position 12 ended up at position 6.
If you look at this list, you can spot a few patterns. First, notice that the final index of all the even-numbered elements is always half the position of that element. For example, the element at position 4 ended up at position 2, the element at position 12 ended up at position 6, etc. This makes sense - we pushed all the even elements to the front of the array, so half of the elements that came before them will have been displaced and moved out of the way.
Now, what about the odd-numbered elements? Well, there are 12 total elements. Each odd-numbered element gets pushed to the second half, so an odd-numbered element at position 2k+1 will get pushed to at least position 7. Its position within the second half is given by the value of k. Therefore, the elements at an odd position 2k+1 gets mapped to position 7 + k.
We can take a minute to generalize this idea. Suppose that the array we're permuting has length 2n. An element at position 2x will be mapped to position x (again, even numbers get halfed), and an element at position 2x+1 will be mapped to position n + 1 + x. Restating this:
The final position of an element at position p is determined as follows:
If p = 2x for some integer x, then 2x ↦ x
If p = 2x+1 for some integer x, then 2x+1 ↦ n + 1 + x
And now we're going to do something that's entirely crazy and unexpected. Right now, we have a piecewise rule for determining where each element ends up: we either divide by two, or we do something weird involving n + 1. However, from a number-theoretic perspective, there is a single, unified rule explaining where all elements are supposed to end up.
The insight we need is that in both cases, it seems like, in some way, we're dividing the index by two. For the even case, the new index really is formed by just dividing by two. For the odd case, the new index kinda looks like it's formed by dividing by two (notice that 2x+1 went to x + (n + 1)), but there's an extra term in there. In a number-theoretic sense, though, both of these really correspond to division by two. Here's why.
Rather than taking the source index and dividing by two to get the destination index, what if we take the destination index and multiply by two? If we do that, an interesting pattern emerges.
Suppose our original number was 2x. The destination is then x, and if we double the destination index to get back 2x, we end up with the source index.
Now suppose that our original number was 2x+1. The destination is then n + 1 + x. Now, what happens if we double the destination index? If we do that, we get back 2n + 2 + 2x. If we rearrange this, we can alternatively rewrite this as (2x+1) + (2n+1). In other words, we've gotten back the original index, plus an extra (2n+1) term.
Now for the kicker: what if all of our arithmetic is done modulo 2n + 1? In that case, if our original number was 2x + 1, then twice the destination index is (2x+1) + (2n+1) = 2x + 1 (modulo 2n+1). In other words, the destination index really is half of the source index, just done modulo 2n+1!
This leads us to a very, very interesting insight: the ultimate destination of each of the elements in a 2n-element array is given by dividing that number by two, modulo 2n+1. This means that there really is a nice, unified rule for determining where everything goes. We just need to be able to divide by two modulo 2n+1. It just happens to work out that in the even case, this is normal integer division, and in the odd case, it works out to taking the form n + 1 + x.
Consequently, we can reframe our problem in the following way: given a 1-indexed array of 2n elements, how do we permute the elements so that each element that was originally at index x ends up at position x/2 mod (2n+1)?
Cycle Decompositions Revisited
At this point, we've made quite a lot of progress. Given any element, we know where that element should end up. If we can figure out a nice way to get a cycle decomposition of the overall permutation, we're done.
This is, unfortunately, where things get complicated. Suppose, for example, that our array has 10 elements. In that case, we want to transform the array like this:
Initial: 1 2 3 4 5 6 7 8 9 10
Final: 2 4 6 8 10 1 3 5 7 9
The cycle decomposition of this permutation is (1 6 3 7 9 10 5 8 4 2). If our array has 12 elements, we want to transform it like this:
Initial: 1 2 3 4 5 6 7 8 9 10 11 12
Final: 2 4 6 8 10 12 1 3 5 7 9 11
This has cycle decomposition (1 7 10 5 9 11 12 6 3 8 4 2 1). If our array has 14 elements, we want to transform it like this:
Initial: 1 2 3 4 5 6 7 8 9 10 11 12 13 14
Final: 2 4 6 8 10 12 14 1 3 5 7 9 11 13
This has cycle decomposition (1 8 4 2)(3 9 12 6)(5 10)(7 11 13 14). If our array has 16 elements, we want to transform it like this:
Initial: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Final: 2 4 6 8 10 12 14 16 1 3 5 7 9 11 13 15
This has cycle decomposition (1 9 13 15 16 8 4 2)(3 10 5 11 14 7 12 6).
The problem here is that these cycles don't seem to follow any predictable patterns. This is a real problem if we're going to try to solve this problem in O(1) space and O(n) time. Even though given any individual element we can figure out what cycle contains it and we can efficiently shuffle that cycle, it's not clear how we figure out what elements belong to what cycles, how many different cycles there are, etc.
Primitive Roots
This is where number theory comes in. Remember that each element's new position is formed by dividing that number by two, modulo 2n+1. Thinking about this backwards, we can figure out which number will take the place of each number by multiplying by two modulo 2n+1. Therefore, we can think of this problem by finding the cycle decomposition in reverse: we pick a number, keep multiplying it by two and modding by 2n+1, and repeat until we're done with the cycle.
This gives rise to a well-studied problem. Suppose that we start with the number k and think about the sequence k, 2k, 22k, 23k, 24k, etc., all done modulo 2n+1. Doing this gives different patterns depending on what odd number 2n+1 you're modding by. This explains why the above cycle patterns seem somewhat arbitrary.
I have no idea how anyone figured this out, but it turns out that there's a beautiful result from number theory that talks about what happens if you take this pattern mod 3k for some number k:
Theorem: Consider the sequence 3s, 3s·2, 3s·22, 3s·23, 3s·24, etc. all modulo 3k for some k ≥ s. This sequence cycles through through every number between 1 and 3k, inclusive, that is divisible by 3s but not divisible by 3s+1.
We can try this out on a few examples. Let's work modulo 27 = 32. The theorem says that if we look at 3, 3 · 2, 3 · 4, etc. all modulo 27, then we should see all the numbers less than 27 that are divisible by 3 and not divisible by 9. Well, let'see what we get:
3 · 20 = 3 · 1 = 3 = 3 mod 27
3 · 21 = 3 · 2 = 6 = 6 mod 27
3 · 22 = 3 · 4 = 12 = 12 mod 27
3 · 23 = 3 · 8 = 24 = 24 mod 27
3 · 24 = 3 · 16 = 48 = 21 mod 27
3 · 25 = 3 · 32 = 96 = 15 mod 27
3 · 26 = 3 · 64 = 192 = 3 mod 27
We ended up seeing 3, 6, 12, 15, 21, and 24 (though not in that order), which are indeed all the numbers less than 27 that are divisible by 3 but not divisible by 9.
We can also try this working mod 27 and considering 1, 2, 22, 23, 24 mod 27, and we should see all the numbers less than 27 that are divisible by 1 and not divisible by 3. In other words, this should give back all the numbers less than 27 that aren't divisible by 3. Let's see if that's true:
20 = 1 = 1 mod 27
21 = 2 = 2 mod 27
22 = 4 = 4 mod 27
23 = 8 = 8 mod 27
24 = 16 = 16 mod 27
25 = 32 = 5 mod 27
26 = 64 = 10 mod 27
27 = 128 = 20 mod 27
28 = 256 = 13 mod 27
29 = 512 = 26 mod 27
210 = 1024 = 25 mod 27
211 = 2048 = 23 mod 27
212 = 4096 = 19 mod 27
213 = 8192 = 11 mod 27
214 = 16384 = 22 mod 27
215 = 32768 = 17 mod 27
216 = 65536 = 7 mod 27
217 = 131072 = 14 mod 27
218 = 262144 = 1 mod 27
Sorting these, we got back the numbers 1, 2, 4, 5, 7, 8, 10, 11, 13, 14, 16, 17, 19, 20, 22, 23, 25, 26 (though not in that order). These are exactly the numbers between 1 and 26 that aren't multiples of three!
This theorem is crucial to the algorithm for the following reason: if 2n+1 = 3k for some number k, then if we process the cycle containing 1, it will properly shuffle all numbers that aren't multiples of three. If we then start the cycle at 3, it will properly shuffle all numbers that are divisible by 3 but not by 9. If we then start the cycle at 9, it will properly shuffle all numbers that are divisible by 9 but not by 27. More generally, if we use the cycle shuffle algorithm on the numbers 1, 3, 9, 27, 81, etc., then we will properly reposition all the elements in the array exactly once and will not have to worry that we missed anything.
So how does this connect to 3k + 1? Well, we need to have that 2n + 1 = 3k, so we need to have that 2n = 3k - 1. But remember - we dropped the very first and very last element of the array when we did this! Adding those back in tells us that we need blocks of size 3k + 1 for this procedure to work correctly. If the blocks are this size, then we know for certain that the cycle decomposition will consist of a cycle containing 1, a nonoverlapping cycle containing 3, a nonoverlapping cycle containing 9, etc. and that these cycles will contain all the elements of the array. Consequently, we can just start cycling 1, 3, 9, 27, etc. and be absolutely guaranteed that everything gets shuffled around correctly. That's amazing!
And why is this theorem true? It turns out that a number k for which 1, k, k2, k3, etc. mod pn that cycles through all the numbers that aren't multiples of p (assuming p is prime) is called a primitive root of the number pn. There's a theorem that says that 2 is a primitive root of 3k for all numbers k, which is why this trick works. If I have time, I'd like to come back and edit this answer to include a proof of this result, though unfortunately my number theory isn't at a level where I know how to do this.
Summary
This problem was tons of fun to work on. It involves cute tricks with dividing by two modulo an odd numbers, cycle decompositions, primitive roots, and powers of three. I'm indebted to this arXiv paper which described a similar (though quite different) algorithm and gave me a sense for the key trick behind the technique, which then let me work out the details for the algorithm you described.
Hope this helps!
Here is most of the mathematical argument missing from templatetypedef’s
answer. (The rest is comparatively boring.)
Lemma: for all integers k >= 1, we have
2^(2*3^(k-1)) = 1 + 3^k mod 3^(k+1).
Proof: by induction on k.
Base case (k = 1): we have 2^(2*3^(1-1)) = 4 = 1 + 3^1 mod 3^(1+1).
Inductive case (k >= 2): if 2^(2*3^(k-2)) = 1 + 3^(k-1) mod 3^k,
then q = (2^(2*3^(k-2)) - (1 + 3^(k-1)))/3^k.
2^(2*3^(k-1)) = (2^(2*3^(k-2)))^3
= (1 + 3^(k-1) + 3^k*q)^3
= 1 + 3*(3^(k-1)) + 3*(3^(k-1))^2 + (3^(k-1))^3
+ 3*(1+3^(k-1))^2*(3^k*q) + 3*(1+3^(k-1))*(3^k*q)^2 + (3^k*q)^3
= 1 + 3^k mod 3^(k+1).
Theorem: for all integers i >= 0 and k >= 1, we have
2^i = 1 mod 3^k if and only if i = 0 mod 2*3^(k-1).
Proof: the “if” direction follows from the Lemma. If
i = 0 mod 2*3^(k-1), then
2^i = (2^(2*3^(k-1)))^(i/(2*3^(k-1)))
= (1+3^k)^(i/(2*3^(k-1))) mod 3^(k+1)
= 1 mod 3^k.
The “only if” direction is by induction on k.
Base case (k = 1): if i != 0 mod 2, then i = 1 mod 2, and
2^i = (2^2)^((i-1)/2)*2
= 4^((i-1)/2)*2
= 2 mod 3
!= 1 mod 3.
Inductive case (k >= 2): if 2^i = 1 mod 3^k, then
2^i = 1 mod 3^(k-1), and the inductive hypothesis implies that
i = 0 mod 2*3^(k-2). Let j = i/(2*3^(k-2)). By the Lemma,
1 = 2^i mod 3^k
= (1+3^(k-1))^j mod 3^k
= 1 + j*3^(k-1) mod 3^k,
where the dropped terms are divisible by (3^(k-1))^2, so
j = 0 mod 3, and i = 0 mod 2*3^(k-1).
given a sorted array of distinct integers, what is the minimum number of steps required to make the integers contiguous? Here the condition is that: in a step , only one element can be changed and can be either increased or decreased by 1 . For example, if we have 2,4,5,6 then '2' can be made '3' thus making the elements contiguous(3,4,5,6) .Hence the minimum steps here is 1 . Similarly for the array: 2,4,5,8:
Step 1: '2' can be made '3'
Step 2: '8' can be made '7'
Step 3: '7' can be made '6'
Thus the sequence now is 3,4,5,6 and the number of steps is 3.
I tried as follows but am not sure if its correct?
//n is the number of elements in array a
int count=a[n-1]-a[0]-1;
for(i=1;i<=n-2;i++)
{
count--;
}
printf("%d\n",count);
Thanks.
The intuitive guess is that the "center" of the optimal sequence will be the arithmetic average, but this is not the case. Let's find the correct solution with some vector math:
Part 1: Assuming the first number is to be left alone (we'll deal with this assumption later), calculate the differences, so 1 12 3 14 5 16-1 2 3 4 5 6 would yield 0 -10 0 -10 0 -10.
sidenote: Notice that a "contiguous" array by your implied definition would be an increasing arithmetic sequence with difference 1. (Note that there are other reasonable interpretations of your question: some people may consider 5 4 3 2 1 to be contiguous, or 5 3 1 to be contiguous, or 1 2 3 2 3 to be contiguous. You also did not specify if negative numbers should be treated any differently.)
theorem: The contiguous numbers must lie between the minimum and maximum number. [proof left to reader]
Part 2: Now returning to our example, assuming we took the 30 steps (sum(abs(0 -10 0 -10 0 -10))=30) required to turn 1 12 3 14 5 16 into 1 2 3 4 5 6. This is one correct answer. But 0 -10 0 -10 0 -10+c is also an answer which yields an arithmetic sequence of difference 1, for any constant c. In order to minimize the number of "steps", we must pick an appropriate c. In this case, each time we increase or decrease c, we increase the number of steps by N=6 (the length of the vector). So for example if we wanted to turn our original sequence 1 12 3 14 5 16 into 3 4 5 6 7 8 (c=2), then the differences would have been 2 -8 2 -8 2 -8, and sum(abs(2 -8 2 -8 2 -8))=30.
Now this is very clear if you could picture it visually, but it's sort of hard to type out in text. First we took our difference vector. Imagine you drew it like so:
4|
3| *
2| * |
1| | | *
0+--+--+--+--+--*
-1| |
-2| *
We are free to "shift" this vector up and down by adding or subtracting 1 from everything. (This is equivalent to finding c.) We wish to find the shift which minimizes the number of | you see (the area between the curve and the x-axis). This is NOT the average (that would be minimizing the standard deviation or RMS error, not the absolute error). To find the minimizing c, let's think of this as a function and consider its derivative. If the differences are all far away from the x-axis (we're trying to make 101 112 103 114 105 116), it makes sense to just not add this extra stuff, so we shift the function down towards the x-axis. Each time we decrease c, we improve the solution by 6. Now suppose that one of the *s passes the x axis. Each time we decrease c, we improve the solution by 5-1=4 (we save 5 steps of work, but have to do 1 extra step of work for the * below the x-axis). Eventually when HALF the *s are past the x-axis, we can NO LONGER IMPROVE THE SOLUTION (derivative: 3-3=0). (In fact soon we begin to make the solution worse, and can never make it better again. Not only have we found the minimum of this function, but we can see it is a global minimum.)
Thus the solution is as follows: Pretend the first number is in place. Calculate the vector of differences. Minimize the sum of the absolute value of this vector; do this by finding the median OF THE DIFFERENCES and subtracting that off from the differences to obtain an improved differences-vector. The sum of the absolute value of the "improved" vector is your answer. This is O(N) The solutions of equal optimality will (as per the above) always be "adjacent". A unique solution exists only if there are an odd number of numbers; otherwise if there are an even number of numbers, AND the median-of-differences is not an integer, the equally-optimal solutions will have difference-vectors with corrective factors of any number between the two medians.
So I guess this wouldn't be complete without a final example.
input: 2 3 4 10 14 14 15 100
difference vector: 2 3 4 5 6 7 8 9-2 3 4 10 14 14 15 100 = 0 0 0 -5 -8 -7 -7 -91
note that the medians of the difference-vector are not in the middle anymore, we need to perform an O(N) median-finding algorithm to extract them...
medians of difference-vector are -5 and -7
let us take -5 to be our correction factor (any number between the medians, such as -6 or -7, would also be a valid choice)
thus our new goal is 2 3 4 5 6 7 8 9+5=7 8 9 10 11 12 13 14, and the new differences are 5 5 5 0 -3 -2 -2 -86*
this means we will need to do 5+5+5+0+3+2+2+86=108 steps
*(we obtain this by repeating step 2 with our new target, or by adding 5 to each number of the previous difference... but since you only care about the sum, we'd just add 8*5 (vector length times correct factor) to the previously calculated sum)
Alternatively, we could have also taken -6 or -7 to be our correction factor. Let's say we took -7...
then the new goal would have been 2 3 4 5 6 7 8 9+7=9 10 11 12 13 14 15 16, and the new differences would have been 7 7 7 2 1 0 0 -84
this would have meant we'd need to do 7+7+7+2+1+0+0+84=108 steps, the same as above
If you simulate this yourself, can see the number of steps becomes >108 as we take offsets further away from the range [-5,-7].
Pseudocode:
def minSteps(array A of size N):
A' = [0,1,...,N-1]
diffs = A'-A
medianOfDiffs = leftMedian(diffs)
return sum(abs(diffs-medianOfDiffs))
Python:
leftMedian = lambda x:sorted(x)[len(x)//2]
def minSteps(array):
target = range(len(array))
diffs = [t-a for t,a in zip(target,array)]
medianOfDiffs = leftMedian(diffs)
return sum(abs(d-medianOfDiffs) for d in diffs)
edit:
It turns out that for arrays of distinct integers, this is equivalent to a simpler solution: picking one of the (up to 2) medians, assuming it doesn't move, and moving other numbers accordingly. This simpler method often gives incorrect answers if you have any duplicates, but the OP didn't ask that, so that would be a simpler and more elegant solution. Additionally we can use the proof I've given in this solution to justify the "assume the median doesn't move" solution as follows: the corrective factor will always be in the center of the array (i.e. the median of the differences will be from the median of the numbers). Thus any restriction which also guarantees this can be used to create variations of this brainteaser.
Get one of the medians of all the numbers. As the numbers are already sorted, this shouldn't be a big deal. Assume that median does not move. Then compute the total cost of moving all the numbers accordingly. This should give the answer.
community edit:
def minSteps(a):
"""INPUT: list of sorted unique integers"""
oneMedian = a[floor(n/2)]
aTarget = [oneMedian + (i-floor(n/2)) for i in range(len(a))]
# aTargets looks roughly like [m-n/2?, ..., m-1, m, m+1, ..., m+n/2]
return sum(abs(aTarget[i]-a[i]) for i in range(len(a)))
This is probably not an ideal solution, but a first idea.
Given a sorted sequence [x1, x2, …, xn]:
Write a function that returns the differences of an element to the previous and to the next element, i.e. (xn – xn–1, xn+1 – xn).
If the difference to the previous element is > 1, you would have to increase all previous elements by xn – xn–1 – 1. That is, the number of necessary steps would increase by the number of previous elements × (xn – xn–1 – 1). Let's call this number a.
If the difference to the next element is >1, you would have to decrease all subsequent elements by xn+1 – xn – 1. That is, the number of necessary steps would increase by the number of subsequent elements × (xn+1 – xn – 1). Let's call this number b.
If a < b, then increase all previous elements until they are contiguous to the current element. If a > b, then decrease all subsequent elements until they are contiguous to the current element. If a = b, it doesn't matter which of these two actions is chosen.
Add up the number of steps taken in the previous step (by increasing the total number of necessary steps by either a or b), and repeat until all elements are contiguous.
First of all, imagine that we pick an arbitrary target of contiguous increasing values and then calculate the cost (number of steps required) for modifying the array the array to match.
Original: 3 5 7 8 10 16
Target: 4 5 6 7 8 9
Difference: +1 0 -1 -1 -2 -7 -> Cost = 12
Sign: + 0 - - - -
Because the input array is already ordered and distinct, it is strictly increasing. Because of this, it can be shown that the differences will always be non-increasing.
If we change the target by increasing it by 1, the cost will change. Each position in which the difference is currently positive or zero will incur an increase in cost by 1. Each position in which the difference is currently negative will yield a decrease in cost by 1:
Original: 3 5 7 8 10 16
New target: 5 6 7 8 9 10
New Difference: +2 +1 0 0 -1 -6 -> Cost = 10 (decrease by 2)
Conversely, if we decrease the target by 1, each position in which the difference is currently positive will yield a decrease in cost by 1, while each position in which the difference is zero or negative will incur an increase in cost by 1:
Original: 3 5 7 8 10 16
New target: 3 4 5 6 7 8
New Difference: 0 -1 -2 -2 -3 -8 -> Cost = 16 (increase by 4)
In order to find the optimal values for the target array, we must find a target such that any change (increment or decrement) will not decrease the cost. Note that an increment of the target can only decrease the cost when there are more positions with negative difference than there are with zero or positive difference. A decrement can only decrease the cost when there are more positions with a positive difference than with a zero or negative difference.
Here are some example distributions of difference signs. Remember that the differences array is non-increasing, so positives always have to be first and negatives last:
C C
+ + + - - - optimal
+ + 0 - - - optimal
0 0 0 - - - optimal
+ 0 - - - - can increment (negatives exceed positives & zeroes)
+ + + 0 0 0 optimal
+ + + + - - can decrement (positives exceed negatives & zeroes)
+ + 0 0 - - optimal
+ 0 0 0 0 0 optimal
C C
Observe that if one of the central elements (marked C) is zero, the target must be optimal. In such a circumstance, at best any increment or decrement will not change the cost, but it may increase it. This result is important, because it gives us a trivial solution. We pick a target such that a[n/2] remains unchanged. There may be other possible targets that yield the same cost, but there are definitely none that are better. Here's the original code modified to calculate this cost:
//n is the number of elements in array a
int targetValue;
int cost = 0;
int middle = n / 2;
int startValue = a[middle] - middle;
for (i = 0; i < n; i++)
{
targetValue = startValue + i;
cost += abs(targetValue - a[i]);
}
printf("%d\n",cost);
You can not do it by iterating once on the array, that's for sure.
You need first to check the difference between each two numbers, for example:
2,7,8,9 can be 2,3,4,5 with 18 steps or 6,7,8,9 with 4 steps.
Create a new array with the difference like so: for 2,7,8,9 it wiil be 4,1,1. Now you can decide whether to increase or decrease the first number.
Lets assume that the contiguous array looks something like this -
c c+1 c+2 c+3 .. and so on
Now lets take an example -
5 7 8 10
The contiguous array in this case will be -
c c+1 c+2 c+3
In order to get the minimum steps, the sum of the modulus of the difference of the integers(before and after) w.r.t the ith index should be the minimum. In which case,
(c-5)^2 + (c-6)^2 + (c-6)^2 + (c-7)^2 should be minimum
Let f(c) = (c-5)^2 + (c-6)^2 + (c-6)^2 + (c-7)^2
= 4c^2 - 48c + 146
Applying differential calculus to get the minima,
f'(c) = 8c - 48 = 0
=> c = 6
So our contiguous array is 6 7 8 9 and the minimum cost here is 2.
To sum it up, just generate f(c), get the first differential and find out c.
This should take O(n).
Brute force approach O(N*M)
If one draws a line through each point in the array a then y0 is a value where each line starts at index 0. Then the answer is the minimum among number of steps reqired to get from a to every line that starts at y0, in Python:
y0s = set((y - i) for i, y in enumerate(a))
nsteps = min(sum(abs(y-(y0+i)) for i, y in enumerate(a))
for y0 in xrange(min(y0s), max(y0s)+1)))
Input
2,4,5,6
2,4,5,8
Output
1
3
This question already has answers here:
Expand a random range from 1–5 to 1–7
(78 answers)
Closed 8 years ago.
How can I generate a bigger probability set from a smaller probability set?
This is from Algorithm Design Manual -Steven Skiena
Q:
Use a random number generator (rng04) that generates numbers from {0,1,2,3,4} with equal probability to write a random number generator that generates numbers from 0 to 7 (rng07) with equal probability?
I tried for around 3 hours now, mostly based on summing two rng04 outputs. The problem is that in that case the probability of each value is different - 4 can come with 5/24 probability while 0 happening is 1/24. I tried some ways to mask it, but cannot.
Can somebody solve this?
You have to find a way to combine the two sets of random numbers (the first and second random {0,1,2,3,4} ) and make n*n distinct possibilities. Basically the problem is that with addition you get something like this
X
0 1 2 3 4
0 0 1 2 3 4
Y 1 1 2 3 4 5
2 2 3 4 5 6
3 3 4 5 6 7
4 4 5 6 7 8
Which has duplicates, which is not what you want. One possible way to combine the two sets would be the Z = X + Y*5 where X and Y are the two random numbers. That would give you a set of results like this
X
0 1 2 3 4
0 0 1 2 3 4
Y 1 5 6 7 8 9
2 10 11 12 13 14
3 15 16 17 18 19
4 20 21 22 23 24
So now that you have a bigger set of random numbers, you need to do the reverse and make it smaller. This set has 25 distinct values (because you started with 5, and used two random numbers, so 5*5=25). The set you want has 8 distinct values. A naïve way to do this would be
x = rnd(5) // {0,1,2,3,4}
y = rnd(5) // {0,1,2,3,4}
z = x+y*5 // {0-24}
random07 = x mod 8
This would indeed have a range of {0,7}. But the values {1,7} would appear 3/25 times, and the value 0 would appear 4/25 times. This is because 0 mod 8 = 0, 8 mod 8 = 0, 16 mod 8 = 0 and 24 mod 8 = 0.
To fix this, you can modify the code above to this.
do {
x = rnd(5) // {0,1,2,3,4}
y = rnd(5) // {0,1,2,3,4}
z = x+y*5 // {0-24}
while (z != 24)
random07 = z mod 8
This will take the one value (24) that is throwing off your probabilities and discard it. Generating a new random number if you get a 'bad' value like this will make your algorithm run very slightly longer (in this case 1/25 of the time it will take 2x as long to run, 1/625 it will take 3x as long, etc). But it will give you the right probabilities.
The real problem, of course, is the fact that the numbers in the middle of the sum (4 in this case) occur in many combinations (0+4, 1+3, etc.) whereas 0 and 8 have exactly one way to be produced.
I don't know how to solve this problem, but I'm going to try to reduce it a bit for you. Some points to consider:
The 0-7 range has 8 possible values, so ultimately the total number of possible situations that you should aim for has to be a multiple of 8. That way you can have an integral number of distributions per value in that codomain.
When you take the sum of two density functions, the number of possible situations (not necessarily distinct when you evaluate the sum, just in terms of different permutations of inputs) is equal to the product of the size of each of the input sets.
Thus, given two {0,1,2,3,4} sets summed together, you have 5*5=25 possibilities.
It will not be possible to get a multiple of eight (see first point) from powers of 5 (see second point, but extrapolate it to any number of sets > 1), so you will need to have a surplus of possible situations in your function and ignore some of them if they occur.
The simplest way to do that, as far as I can see at this point, is to use the sum of two {0,1,2,3,4} sets (25 possibilities) and ignore 1 (to leave 24, a multiple of 8).
Thus the challenge now has been reduced to this: Find a way to distribute the remaining 24 possibilities among the 8 output values. For this, you'll probably NOT want to use the sum, but rather just the input values.
One way to do that is, imagine a number in base 5 constructed from your input. Ignore 44 (that's your 25th, superfluous value; if you get it, synthesize a new set of inputs) and take the others, modulo 8, and you'll get your 0-7 across 24 different input combinations (3 each), which is an equal distribution.
My logic would be this:
rn07 = 0;
do {
num = rng04;
}
while(num == 4);
rn07 = num * 2;
do {
num = rng04;
}
while(num == 4);
rn07 += num % 2